# Invariant Representations in Deep Learning for Optoacoustic Imaging

**Authors:** Matias Vera, Martin G. Gonzalez, Leonardo Rey Vega  
**arXiv:** [2305.00335](https://arxiv.org/abs/2305.00335)  
**Date:** 2023-04-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.00335v1/

## Key points

- Optoacoustic tomography (OAT) is a field with potential for non-invasive imaging
- Image reconstruction in OAT depends on physical measurements taken at the time of sensing, which can lead to algorithms tailored to specific configurations
- Deep learning algorithms based on learning invariant and robust representations can address this issue
- The authors investigate the use of deep learning algorithms that are robust to different environments or invariant across such environments, using the ANDMask scheme
- Numerical experiments show that when out-of-distribution generalization is imposed against variations in parameters such as sensor location, there is no degradation of performance; in some cases it is even possible to achieve improvements compared to standard deep learning approaches where invariance robustness is not explicitly considered
- Learning reconstruction algorithms that are robust or invariant can be highly valuable because it allows researchers to focus on what truly matters for a given application while discarding spurious features
- The ability to generalize beyond specific configurations could also make OAT more accessible for practical use.
- Deep learning can be leveraged for improved image reconstruction in optoacoustic tomography by considering issues related to variability and uncertainty in physical measurements.

## Summary

The field of optoacoustic tomography (OAT) has seen a surge in interest due to its potential for non-invasive imaging. However, image reconstruction in OAT is highly dependent on physical measurements taken at the time of sensing, and the large number of different settings and uncertainties can lead to reconstruction algorithms that are tailored to specific configurations. This presents a challenge when faced with practical situations where the configuration may differ from what was originally used to design the algorithm. To address this issue, researchers have explored the use of deep learning algorithms based on learning invariant and robust representations for the OAT inverse problem. In their work titled "Invariant Representations in Deep Learning for Optoacoustic Imaging," Matias Vera, Martin G. Gonzalez, and Leonardo Rey Vega investigate the application of deep learning algorithms that are robust to different environments or invariant across such environments. They consider the ANDMask scheme due to its easy adaptation to the OAT problem and conduct numerical experiments showing that when out-of-distribution generalization is imposed against variations in parameters such as sensor location, there is no degradation of performance; in some cases it is even possible to achieve improvements compared to standard deep learning approaches where invariance robustness is not explicitly considered. The authors' findings suggest that learning reconstruction algorithms that are robust or invariant can be highly valuable because it allows researchers to focus on what truly matters for a given application while discarding spurious features. The ability to generalize beyond specific configurations could also make OAT more accessible for practical use. Overall, this work highlights how deep learning can be leveraged for improved image reconstruction in optoacoustic tomography by considering issues related to variability and uncertainty in physical measurements.